Can decentralized teams outperform central planners in long-running science?

Most AI-for-science agents follow a single research trajectory or coordinate through a central planner with fixed objectives. That assumption breaks for long-running experimentation, where research directions are not known in advance and change as evidence arrives. Long-horizon science needs three things short-horizon optimization does not: maintaining competing hypotheses, updating them as evidence shifts, and using failures to redirect the search.

AutoScientists meets these with a decentralized design. Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before consuming experimental compute, and share both successes and failures to reduce redundant exploration. Under matched experimental budgets it beats prior agents across biomedical ML, language-model training optimization, and protein fitness prediction (74.4% mean leaderboard percentile across 24 BioML-Bench tasks, +8.33% over the strongest baseline).

The honest framing matters: AutoScientists is not more LLM-call efficient — it uses more tokens for parallel reasoning, discussion, and team reorganization. Its win is under a fixed experimental-compute budget, by selecting better experiments to run. That is the right efficiency frontier for science, where wet-lab or GPU experiments dominate cost, not inference tokens. This connects to Can experiment failures drive progress instead of stopping it? — AutoScientists makes failure-as-information a team-level shared resource — and to Do self-organizing agent teams outperform rigid hierarchies?, which supplies the coordination evidence for why decentralization beats the central planner.